The g Factor: What General Intelligence Actually Is

The one thing everyone agrees on

In 1904 Charles Spearman published a paper in the American Journal of Psychology noting something odd about schoolchildren's marks. Performance in classics, French, English, mathematics and even pitch discrimination all correlated positively. Subjects with nothing obvious in common tracked one another. Spearman proposed that every test measures two things: a general ability shared by all of them, which he labelled g, and something specific to that particular test, which he labelled s. That is the two-factor theory, and it is where the whole field starts. The observation itself has never been overturned. It is one of the most replicated findings in psychology. Give a large sample any battery of cognitive tasks — verbal, spatial, numerical, speeded, untimed — and the correlation matrix comes back positive, essentially everywhere. This is why IQ tests can produce a single composite score at all. The most thorough audit of that evidence is John Carroll's Human Cognitive Abilities (1993), which reanalysed more than 460 datasets collected over sixty years. Carroll's three-stratum model puts narrow abilities at the bottom, broad abilities such as fluid and crystallized intelligence, memory and processing speed in the middle, and a single general factor at the top. It remains the most complete factor-analytic survey ever assembled. Keep two things apart The positive manifold is the finding. g is the explanation offered for it. Nearly every confusion about general intelligence comes from treating those as the same claim.

What the g factor actually is

g is never measured. It is extracted. You administer a battery of tests, compute the correlations between them, and fit a model in which a latent variable accounts for the variance they share. g is that latent variable. Nobody has ever observed a g; they observe test scores and infer a factor that explains why the scores hang together. This is ordinary psychometric practice, and it is worth being precise about what it does and does not deliver.

• g is a property of a correlation matrix, and a correlation matrix is a property of a population — not of a person.
• The size and even the composition of g depend on which tests you put in the battery and who you administer them to.
• A factor score is an estimate with a standard error, not a reading off a dial.
• Extracting a factor demonstrates that the tests share variance. It does not identify what the shared thing is.

One practical consequence gets lost constantly: your IQ is not your g. An IQ score is a normed composite, scaled so the population mean is 100, and it is read as a percentile against a standardisation sample. g is a latent variable in a factor model. The two are related — IQ is heavily g-loaded — but they are not the same object, and no test hands you your g.

The mistake almost every article makes

Here is the inference you will find nearly everywhere: the tests all correlate, therefore a single underlying thing must be causing them to correlate, therefore general intelligence is a real quantity in the brain. That inference does not follow, and the reason is worth understanding, because it is the single most important thing to know about g. Factor analysis is agnostic about causes. It tells you that a set of variables share variance and gives you a mathematical summary of that sharing. It cannot tell you whether the sharing arises from one common cause, from many causes, or from the variables having influenced each other. Different causal stories can produce the very same correlation matrix — and the data cannot choose between them. This is not a philosophical quibble. In 2006, van der Maas and colleagues published a dynamical model in Psychological Review in which cognitive processes start out uncorrelated and grow by helping each other develop — working memory supporting reasoning, reasoning supporting vocabulary, and so on. They called it mutualism. Run the model, and the positive manifold appears. There is no general factor anywhere in it. The correlations are an emergent consequence of development, not the fingerprint of a common cause. A decade later, Kovacs and Conway proposed Process Overlap Theory in Psychological Inquiry, reaching a similar destination by a different road. On their account what is real are domain-general executive processes — closely tied to working memory and executive function — plus domain-specific processes. Many tests tap the same executive processes, so they correlate. g falls out as a statistical abstraction, not as a causal entity. What this does and does not mean Two formal models, published in the field's leading theory journals, reproduce the positive manifold without any general factor in them. That does not prove g is unreal. It proves that the correlations alone cannot settle the question. Any page telling you factor analysis proves a single underlying intelligence is overclaiming — and so is any page telling you g has been debunked. The honest position is narrower and more useful than either slogan. The positive manifold is beyond dispute. g is an excellent statistical summary of it. Whether g corresponds to one mechanism in the head is genuinely unresolved, and the people who study it hardest are the ones least likely to tell you otherwise.

Why g still matters when you take a real test

None of the above makes g useless. It makes it a statistic rather than an organ — and the statistic turns out to carry most of the weight in the tests clinicians actually use. When Canivez and colleagues reanalysed the structure of the WISC-V in 2017, and later the WAIS-5, they used bifactor models, which let you ask how much of the reliable variance belongs to the general factor and how much is left over for the group factors — verbal comprehension, working memory, processing speed and the rest. The answer, in both cases, was lopsided. The general factor accounted for most of the reliable common variance. The group factors added little that was unique once g was accounted for. That has a direct, practical consequence for anyone reading a score report. The Full Scale IQ — the most g-loaded number on the page — is the part you can interpret with the most confidence. The index profile, the jagged little chart of highs and lows across the four indexes, is far weaker than it looks, because most of what those indexes measure is the same thing the Full Scale already told you. Reading meaning into small peaks and troughs is a standard-error problem before it is anything else. Related: The WAIS test — how the adult Wechsler scale is built, and why independent factor analysis does not support its five-factor structure

g is not a fixed quantity in your head

If g were a single property of an individual brain, it should not much matter who else is in the room. It does matter — demonstrably. Kvist and Gustafsson (2008), writing in Intelligence, tested Cattell's investment theory using Swedish and immigrant groups. The idea behind investment theory is that fluid reasoning gets 'invested' in learning, and crystallized knowledge is the return on that investment. If so, the relationship between fluid ability and g should depend on whether everyone in the sample has had comparable opportunities to learn. That is exactly what they found. In a culturally homogeneous group, fluid intelligence and g were essentially indistinguishable. Pool groups with very different learning opportunities, and the two come apart. Sit with that for a moment. The relationship between fluid reasoning and 'general intelligence' changed depending on the composition of the sample — because g is extracted from a correlation matrix, and the correlation matrix belongs to the group, not the person. Change who is in the group and you change the factor. That is not a flaw in the method. It is what the method is.

What g is not

• Not mental energy. Spearman speculated about a general 'mental energy'. That framing did not survive; nothing in the modern evidence requires it.
• Not a brain region. No structure has been identified as the seat of g. Correlations with brain volume and other measures exist, but they are modest and they are not a location.
• Not your IQ score. IQ is a normed composite; g is a latent variable. Related, not identical.
• Not destiny. g predicts outcomes on average, across groups. What it licenses you to say about one person is far less than most people assume.
• Not a knock-down argument against multiple intelligences — though the honest problem with Gardner's proposal is not that a general factor exists, it is that his proposed intelligences tend to correlate with one another.

And g is not a statement about where ability comes from. Heritability estimates apply to variation within a population under particular conditions, and a high heritability is a fact about a society, not a verdict on an individual. The existence of a general factor tells you nothing on its own about whether ability is fixed. What is left, after all the overclaiming is stripped out, is still substantial: a robust, century-old empirical regularity; a statistical model that summarises it well and predicts real outcomes; and a live, unresolved scientific argument about what sits underneath. That is a more interesting answer than either 'g is the essence of intelligence' or 'g is a myth', and it is the one the evidence actually supports.